Large model at edge: an optimal mobile edge generation (MEG) design

A novel mobile edge generation (MEG) framework is proposed to efficiently operate large models at edge networks for low-latency image generation. The generation of large-scale image content is split into two parts, namely primary and secondary regions, with an adjustable generation splitting ratio. The primary region is generated by a large generative model (LGM) at the edge cloud and then transmitted to the mobile device, whereas the remaining secondary regions is created by a tiny generative model (TinyGM) at the mobile device, thus reducing transmission and computation overheads. Both single-user and multi-user cases are considered to characterize the tradeoff between mobile energy consumption and generation delay. For the single-user case, a multi-objective programming (MOP) is formulated for the joint optimization of generation splitting and mobile power control, which simultaneously minimizes the generation delay and mobile energy consumption. This MOP is transferred into single-objective optimization using the ϵ-constraint method. The closed-form optimal solution is derived to obtain Pareto-optimal energy-delay (E-D) region. It is revealed that MEG achieves significant performance gains then conventional fully edge generation (FEG) when signal-to-noise ratio (SNR) or mobile generative cost is low. For the multi-user case, a joint generation splitting and resource allocation problem is formulated, which minimizes the maximum generation delay subject to ϵ-bounded mobile energy consumption and resource constraints. An McCormick-relaxation branch-and-bound (M-BnB) algorithm is proposed to obtain the globally optimal solution. Simulation results demonstrate the Pareto-optimal E-D region in single-user and multi-user cases. Furthermore, MEG flexibly reduces delay compared to conventional FEG and model split schemes while maintaining generative quality.